1
Utilizing Predictive Analytics
to Increase the Value of Care
Session 237, February 14, 2019
Tina Esposito, Chief Health Information Officer, Advocate Aurora Health
Fran Wilk, Clinical Process Designer, Advocate Aurora Health
2
Tina Esposito, MBA
Fran Wilk, RN, BSN, MA
Have no real or apparent conflicts of interest to report.
Conflict of Interest
3
• Advocate Aurora Health Overview
• Episodic Outpatient Care Management Defined
• Predictive Model Creation
• Workflow Development
• Summary
Agenda
4
• Describe data infrastructure elements needed to develop
predictive models
• Explain how predictive models can be used to identify patients for
clinical programs as well as limitations to this approach
• Develop and apply a prescriptive workflow to achieve a specific
outcome
Learning Objectives
5
• Advocate Aurora Health Overview
• Episodic Outpatient Care Management Defined
• Predictive Model Creation
• Workflow Development
• Summary
Agenda
6
7
7
Contract Lives Total Spend
Commercial HMO
257,000 $0.8B
Medicare
Advantage 47,000 $0.3B
Advocate Employee
32,000 $0.1B
Commercial Shared
Savings
444,000 $1.6B
Medicare Shared Savings
144,000 $1.6B
Medicaid ACO
93,000 $0.1B
Total
~ 1,000,000 $4.5B
AdvocateCare® Population
8
Cost Savings Results
$60.68 million saved
Better than benchmark by 3.9%
Medicare
Shared
Savings Program
10% reduction in medical costs
for EPO vs. PPO
Advocate EPO
8
ACO Successes
9
• Mission
– Leverage Advocate’s experience in clinical integration and
Cerner’s experience in health care automation to improve
population health management capabilities
• Together the collaborative team will innovate to:
– Identify /risk stratify patients at risk
– Facilitate appropriate and early interventions
– Guide care across the continuum
Advocate Cerner Collaborative (ACC)
10
Model
Deployment
Analytic
Models
Data
Platform
Core Competencies: The People
11
Outpatient
Post Acute
Hospital to Home
Transition Coach
Program
SNF Care Model
Palliative Care
Outpatient Care
Managers
Complex Care
Centers
Patient Centered
Medical Home
Chronic Disease
Management
Data & Analytics Population Health Management
Care Coordination
Acute Care
ED Care Coordination
Optimization
Alternative Site Of
Care Transitions
Readmission Risk
Assessment &
Focused
Interventions
Inpatient Care
Coordination
Redesign
Acute To Post Acute
Transitions
ED Care
Coordination
Inpatient Care
Coordination
ED Hospital
AdvocateCare® Programs
12
• Advocate Aurora Health Overview
• Episodic Outpatient Care Management Defined
• Predictive Model Creation
• Workflow Development
• Summary
Agenda
13
14
Intervention Target Population ROI
Acute Case
Management
Hospital < 1 year
Episodic Care
Management
Risk of acute
hospitalization
< 1 year
Disease
Management
Chronic disease
management, e.g.,
Diabetes, Heart
Failure
2-5 years
Complex Care
Management
Multi disease, multi
complication, renal
failure, transplant,
cancer, etc.
2-5 years
Barriers:
• Behavior
• Social
• Adherence
• Education
Barriers:
• Behavior
• Social
• Adherence
• Education
Enablers:
• Care
Management
• Readmission
Prevention
• Transitions of
Care
• Patient Portal
Enablers:
• Care
Management
• Readmission
Prevention
• Transitions of
Care
• Patient Portal
Roles: NP, RN, NA, SW, CHWRoles: NP, RN, NA, SW, CHW
Community Based Care Management
Framework
15
Appropriately identify patients for
short-term care management interventions
• An Effective Program is…
– Designed to reduce ER visits, hospital visits and LOS
– Short term (currently not exceeding 90 days)
– Evidence based
– Measureable
– Focused on potentially preventable events
• Preventable Hospitalizations are:
– Clinician identified preventable events most appropriate for care management.
– Events where Outpatient Care Management (OPCM) intervention can reduce
hospital encounters (ED/IP/OBS) within a 90-day time period
– Impactable in a measurable way, with defined outcomes.
Episodic Care Management Defined
16
• Advocate Aurora Health Overview
• Episodic Outpatient Care Management Defined
• Predictive Model Creation
• Workflow Development
• Summary
Agenda
17
• Who is impactable?
– Patients with conditions like pneumonia, asthma,
enteritis, heart failure, urinary tract infection, COPD, and
dementia/Parkinson’s (7 conditions identified by
clinicians)
• Focus on a short time period
– Develop models that predict the likelihood of an acute
encounter within the next 90 days
Goal: Minimize Acute Encounters
18
What is Impactable?
Condition What is preventable How can we prevent?
Heart
Failure
Acute Exacerbations
ER visits
Non-adherence to
medications
Worsening of symptoms
Education: Heart Failure
Action Plan,
understanding triggers, having a
prevention plan, knowing when and
how to use medications,
understanding consequences of non-
adherence
Adhering to medication treatment
plans
Self
-
management with education and
resources
Early reporting if symptoms appear
19
Category
N
example
Condition
195
Hypertensive_Chronic_Renal_Disease_
cur_yn
Medication
63
M_general_anesthetics_1y_yn
result Lab
46
HEMOGLOBIN_last1y
Symptom
26
Chest_Pain_cur_yn
Utilization
19
observation_1y_CNT
independent_PPE
18
COPD_ppe_1y_yn
result vital
16
BMI_changeCAT
Encounter
Description
9
encounter_typeCat
Demographics
8
race2
dependent_PPE
8
PPE_CHF
SVI
6
avg_R_PL_THEME1
Procedure
3
PNEUMO_1y_yn
cost
1
cost_per_month
Total : 418
• Data extracted
from Hadoop
platform which
aggregates and
normalizes
multiple data
sources
Over 400 Variables Considered
20
Encounter
count
Patient
count
Training
dataset
1,620,516
344,767
Test
dataset
694,152
147,759
total
2,314,668
492,526
Dataset included >2M Encounters
21
model_name
train
AUC
test
AUC
Brier Score
Brier Improve
PPE_Asthma
0.855
0.852
0.005
0.038
PPE_CHF
0.934
0.930
0.006
0.083
PPE_Dehydration
0.833
0.843
0.003
0.011
PPE_Dementia
0.962
0.951
0.002
0.043
PPE_enteritis
0.810
0.790
0.005
0.017
PPE_Hypertension
0.846
0.850
0.013
0.069
Models Have Strong AUC Values
22
0
10
20
30
40
50
60
70
80
0 to 10 10 to 20 20 to 30 30 to 40 40 to 50 50 to 60 60 to 70 70 to 80 80 to 90
COUNT
DAYS SINCE IDENTIFIED
Distribution of Acute Encounters After
Prediction
• Of first 734
patients, >17%
had an acute
encounter
(avg. risk
~24%)
• Acute
encounters
happen early!
Models Performed Well After Launch
23
• Advocate Aurora Health Overview
• Episodic Outpatient Care Management Defined
• Predictive Model Creation
• Workflow Development
• Summary
Agenda
24
• Patients at high risk for going to the hospital in the next 90 days
are not necessarily good candidates for telephonic care
management
• Filters exclude those who may not be a good fit:
– Patients actively engaged with another care team (e.g.
LVAD, heart transplant, hemodialysis, active cancer)
– ADL dependencies
Filtering for OPCM
25
Paradigm Shift
Traditional
Episodic
Generalized
•
Focus on whatever the patient
and care manager agree is the
priority
Disease specific
•
Prioritize Heart Failure with
emphasis on self-management
Custom
•
Collect comprehensive
information on patient and
devise custom care plan
•
Follow general guidelines for
frequency of contact (e.g. once
a month)
Prescriptive
•
Collect information related to
HF and ensure patient
understands and utilizes action
plan
•
Contact frequency is defined
based on patient’s progress
26
Paradigm Shift
Traditional
Episodic
Open
-ended
•
Care manager maintains
involvement if CM and patient
continue to find value in
contacts
•
Inter-Disciplinary Team will
review cases after 6 months
to determine if case should
remain open
Clear Milestones
•
Care manager moves the
patient through well-defined
steps
•
Graduation criteria are well-
defined so CM knows when it
is appropriate to close the
case
27
• Program is
designed to be
prescriptive in its
approach (focus on
disease-specific
action plan)
• Looking to reduce
subjectivity in care
management
process
Prescriptive Workflow
28
• Hire for key attributes:
• Committed to improving patients’ health
• Refined phone etiquette
• Engaging and authentic
• Develop for success:
• Motivational Interviewing
• Teach-back method
• Help them find their own voice
Communication is Central
29
• Monitoring
• All episodic care managers are closely monitored
• Every HF-related acute encounter gets investigated and
discussed
• Project team is always looking for ways to tweak the
program to improve outcomes
• Mentoring
• Care Managers are given lots of coaching and support to
ensure they are successful
• Executive Support
• All levels of leadership are committed to the success of
the program
Other Key Ingredients
30
• Advocate Aurora Health Overview
• Episodic Outpatient Care Management Defined
• Predictive Model Creation
• Workflow Development
• Summary
Agenda
31
• 350+ patients engaged so far
• 50% of all patients who have been closed achieved all
milestones defined in the prescriptive workflow
• Average length of time in program: 70 days (assignment date
to closure date)
• Compared program participants to similar patients who were
not offered this program and observed 23% fewer acute
encounters in the treatment group
Results
32
• A predictive model alone does nothing to keep patients out of
the hospital
• Knowing a patient is at risk of an event doesn’t necessarily
mean you can do anything to prevent it
• Connecting with patients early and often is essential
• Focus on chronic disease self-management first
• Provide care managers with clear objectives and milestones to
ensure consistency across the team
Key Learnings
33
• Adding care managers to ensure we are reaching all HF
patients who will benefit from this program
• Developing predictive models and workflows for additional
diseases (next up: COPD)
• Exploring other interventions that might be able to use these
predictive models to help identify appropriate patients
Next Steps
